Most command methods now take an optional trailing coderef; if a coderef is
supplied, we don't wait for the command's response, but merely schedule the
coderef to be called once the response has been read.
If we aren't in reconnect mode, using __with_reconnect adds a method call
and a closure invocation to every request, no matter how trivial. This
change adds a fast path for directly calling __run_cmd in the three relevant
locations. In my tests, this saves up to a third of the best-case amortised
per-request execution time.